sqlsure

AI writes your SQL. sqlsure makes sure it's right.

A query can be perfectly valid, run without error, and return a number that's silently wrong — revenue double-counted by a join, an average summed, a patient identifier exposed. Databases don't catch this. Linters don't catch this. LLMs reviewing their own SQL don't catch this.

sqlsure does — deterministically, in 0.1 ms, before the query runs.

Proof, not promises: we ran sqlsure over the gold answers of the two benchmarks every text-to-SQL model is graded on. 2,568 expert-written queries, 45 flags, zero false alarms — including a BIRD dev gold answer that is provably wrong by 8× from the exact bug class sqlsure targets, and a schema defect now filed upstream.

How it works

sqlsure judges SQL against facts your team already declared — dbt unique tests become grain, relationships tests become join cardinality, one-line meta tags mark what's safe to sum. No new language to learn, no model to maintain by hand. Rules are dictionary lookups, not LLM calls: same input, same verdict, every time, offline.

Every rejection carries a machine-actionable fix, so AI agents self-repair: draft → check → fix → check → execute. In our benchmark, applying the fix verbatim produced a passing query 10/10 times.

Quick start

pip install sqlsure

from sqlsure import SemanticModel, check
violations = check(sql, model)   # [] means semantically safe

Or clone and run the 30-second demo:

python check.py                   # 5 wrong queries rejected, 1 approved — with fixes
python -m sqlsure.scan path/to/dbt-repo --report report.md   # audit any dbt repo

Three doors, one engine

1. CI gate — blocks the merge when a PR double-counts:

python -m sqlsure.cli --model model.json query.sql   # exit 1 on violations

2. MCP server — your AI agent must pass inspection before executing:

claude mcp add sqlsure -- python -m sqlsure.mcp_server --model /abs/path/model.json

See docs/MCP.md for tool reference and agent-loop patterns.

3. Library — embed check() inside any text-to-SQL product or agent framework. A drop-in SemanticGate wraps Vanna/WrenAI-style generators; a semantic eval metric scores NL2SQL output where execution-accuracy is blind.

Also available as an Agent Skill — a single SKILL.md your agent loads directly; no server process needed.

The rules (v0.1)

Rule	Severity	Catches
FANOUT	error	SUM/COUNT of additive measure after one-to-many join
CHASM	error	two+ fan-out joins multiplying each other
ADDITIVITY	error	SUM of a non-additive measure (rates, averages)
SEMI_ADDITIVE	error	balances/censuses summed across their snapshot dimension
JOIN_KEY	error	join on columns matching no declared relationship
CROSS_JOIN	error	join with no predicate
WEIGHTED_AVG	warning	AVG silently re-weighted by fan-out
UNDECLARED_JOIN	warning	join with no declared relationship (unverifiable ≠ safe)
SENSITIVE_COLUMN	policy	PHI/PII column exposed in query output

When sqlsure can't verify something, it says "can't verify" — never "looks fine." Honest uncertainty is a feature.

Trust properties

• Deterministic — same SQL + same rulebook = same verdict, always; rules are dictionary lookups, auditable line by line
• Offline — zero network calls; your SQL never leaves your machine
• No data access — parses query text; never connects to a database
• No telemetry — nothing collected, ever (SECURITY.md)
• Supply chain — releases ship exclusively via PyPI Trusted Publishing (OIDC) from tagged commits with public CI runs; two runtime deps

Where the rulebook comes from

• dbt (works today): manifest.json or schema.yml — the tests teams already wrote become enforceable semantics, zero config

• Plain PK/FK declarations (works today — powered the benchmark audits)

• The live database itself (works today): no semantic layer at all? sqlsure.introspect builds the rulebook from the catalog — SQLite PRAGMAs or information_schema PK/FK (postgres/mysql). Introspecting BIRD's own database files recovered 2 foreign keys missing from the benchmark's published schema (bird-bench/mini_dev#37)

  from sqlsure.introspect import model_from_sqlite
  model = model_from_sqlite("app.db")   # PK -> grain, FK -> join edges

• Hand-written JSON — model.example.json

• OSI and WrenAI MDL (working loaders in integrations/): OSI demonstrated on the spec's published examples; WrenAI MDL demonstrated on WrenAI's own shipped example manifest — primaryKey → grain, relationship joinType + condition → join edges, cube measures → additivity

• Cube, Snowflake Semantic Views — adapters on the roadmap; the engine only ever sees one SemanticModel

Validated on

• 16/16 rule tests, 100% recall / 0% false positives on the paired benchmark (docs/METRICS.md)
• Real production repos (Mattermost's warehouse, Fivetran packages, dbt's jaffle shop) — docs/TEST-REPORTS.md
• Spider + BIRD gold queries — the zero-noise external audit above

Learn more

• docs/EVIDENCE.md — what it does for you, every claim linked to a rerunnable measurement
• docs/ARCHITECTURE.md — how it physically works, ELI5 → god level, with real intermediate outputs
• docs/FOR-DUMMIES.md — every concept from zero
• docs/INTEGRATIONS.md — GitHub Action, pre-commit, MCP, Snowflake UDF / Cortex Agent tool, query-history audit
• docs/MCP.md — MCP server documentation
• CONTRIBUTING.md — adding rules and loaders

Apache-2.0 · sqlsure.ai

mcp-name: io.github.sqlsure/sqlsure